IBM Watson Shows Us How Machine Learning Is Transforming Cancer Care
Famed for its winning appearance on the popular game show Jeopardy!, IBM Watson has become an important tool for doctors working with cancer patients around the world. Having trained the supercomputer to understand the complexities and nuances of human language, IBM developed Watson for Oncology, which provides information to physicians to help them identify personalized, evidence-based cancer care options. Based on research done in India, it was found that Watson’s treatment recommendations matched those of doctors at a rate of 96 percent for lung, 81 percent for colon and 93 percent for rectal cancer cases. Watson Clinical Trials Matching, meanwhile, evaluates data from patient records and doctor’s notes to automatically find eligible patients for clinical trials. In one Arkansas-based study, this reduced clinical trial screening time from 1 hour and 50 minutes to just 24 minutes.
In our latest Future of Health report, PSFK spoke with Andrew Norden, deputy chief health officer for oncology and genomics at IBM Watson Health, to discuss both current and potential future uses for Watson.
Many of us are familiar with IBM Watson from its appearance onJeopardy!, but could you explain how it’s being used in healthcare?
Healthcare was actually conceived as a critical first-use case for Watson back when IBM Watson was being used on Jeopardy!. IBM opted to partner with Memorial Sloan Kettering in New York City around cancer, and created Watson for Oncology, a tool that helps physicians efficiently and accurately choose an optimal evidence-based therapeutic option for a given patient. Watson for Oncology leverages Watson’s natural language processing capabilities, which we had developed for Jeopardy!, with expert training over the years, so that Watson could read not just common language, but also healthcare language, and specifically oncology language. Watson reads the medical records for a given patient, looking both at the structured and unstructured data. Then once it’s identified the key attributes for a given patient, Watson looks into a corpus of data that includes guidelines, textbooks, thousands of pages of journal articles, and published abstracts, while also relying on its expert training from doctors at Memorial Sloan Kettering over thousands of cases. Then it makes evidence-based treatment recommendations for the physician to consider in deciding how to treat a given patient.
Could you talk more about structured and unstructured data?
In healthcare, a lot of the most important information about a given patient exists in an unstructured narrative form in the electronic medical records. Doctors’ notes are highly variable from one to the next. When you look at those narrative documents, you can’t necessarily predict where in the given document a crucial piece of information might reside. Watson works by reading the text, understanding the content, making conceptual links between sentences and paragraphs, and ultimately is able to extract from that unstructured text the most important pieces of information. It’s able to do that from doctors’ notes, radiology reports, oncology reports, laboratory results, etc. It then compares that against these terrific knowledge sources that are constantly being updated.
How does Watson Clinical Trials Matching make finding matches easier?
There are a lot of barriers to clinical trial enrollment. Every clinical trial in which a cancer patient might participate carries with it a long and detailed list of eligibility criteria that must be met in order for a patient to be enrolled. The eligibility criteria includes things like the cancer type and stage, other co-morbid conditions, and medications that the patient might be on prior to therapy. A given physician who’s enrolling patients in clinical trials typically has tens or even hundreds of trials to consider. The lists of eligibility criteria are generally tens of items long, so that’s obviously a hard thing for any human to keep track of. Watson Clinical Trials Matching benefits physicians and patients in much the same way Watson for Oncology does. It reads the medical records, structured and unstructured data, and pulls out key attributes. It can then, in an automated fashion, compare the patient attributes to the eligibility criteria and help physicians make matches.
How is Watson offering increased personalization to doctors and patients?
I look at Watson as an engine for personalized medicine therapy because its first step is to extract all of the known, relevant attributes of a given patient that need to be considered in making treatment recommendations. As one looks to the future, I think there’s also the reality that there are many likely undiscovered attributes of given patients and given tumors that will impact treatment. I believe that Watson has the capacity to help us identify what some of these factors are.
Watson Oncology has a high concordance with doctors when it comes to treatment recommendations—above 80 or even 90 percent in most studies. Is the goal eventually to get to 100 percent concordance?
This 80, 90 percent that you cite, it is exactly what you say, the concordance between Watson and expert human physicians. But the studies were not intended to answer the question of, “Which is right?” In other words, if you get an 80 percent concordance number, it may be equally probable that Watson is right or the humans are right. If you take a group of well-trained doctors, even working in one hospital, give them a case, and ask for their treatment recommendations, I think, in many cases, you would find quite a low concordance rate between them. That doesn’t necessarily mean that those doctors are right or wrong. Sometimes, it means that there are multiple good answers.
How do you see cognitive computing developing in the future for healthcare? Do you see Watson eventually having the capacity to replace doctors in any way?
I don’t expect that Watson is going to replace doctors. That’s definitely not IBM’s intent, and I don’t think it’s realistic. I do think that cognitive computing in healthcare is going to become increasingly ubiquitous because of the reality that there is such a growing cognitive burden based on the rapid acceleration of knowledge in healthcare. I think that there are some things that cognitive computing systems do extraordinarily well. They can find patterns in large data sets that humans can’t find. They can search very, very quickly, they don’t need to sleep, they have unlimited storage, they’re often not subject to the same kind of cognitive biases that people are—that sort of thing. That said, there is and always will be, I think, a critical role for physicians to sit and look at patients eye-to-eye, to help patients weigh options, to be a compassionate voice, to help people explore their goals, fears, and hopes. The offloading of some mundane or routine tasks from physicians by cognitive computing systems will free them up to be more available for those strictly human types of interactions.
Watson Oncology focuses on treatment recommendations, but are there any other potential uses for Watson in the realm of healthcare?
We’re looking at a wide variety of different use cases for cognitive computing across multiple different areas of healthcare. It’s true that’s today’s Watson Oncology model brings our focus substantially on the provision of treatment recommendations for patients with diagnosed cancer. There’s no doubt that cognitive computing capabilities can have a role in, for example, prediction of disease. Patient subgroups who are at particularly high risk for an illness who might benefit from additional testing, or even prevention strategies. One of our major areas of focus that overlaps with oncology is in imaging. Watson Health Imaging is developing tools to improve the accuracy rate of tomography, for example. Way too many mammogram readings result in false positive or false negative results. Cognitive computing technology is typically the ability of Watson to interpret medical images based on training. We think it has real potential to improve the accuracy of these imaging-based screening studies.
Many of us are familiar with IBM Watson from its appearance onJeopardy!, but could you explain how it’s being used in healthcare?
Healthcare was actually conceived as a critical first-use case for Watson back when IBM Watson was being used on Jeopardy!. IBM opted to partner with Memorial Sloan Kettering in New York City around cancer, and created Watson for Oncology, a tool that helps physicians efficiently and accurately choose an optimal evidence-based therapeutic option for a given patient. Watson for Oncology leverages Watson’s natural language processing capabilities, which we had developed for Jeopardy!, with expert training over the years, so that Watson could read not just common language, but also healthcare language, and specifically oncology language. Watson reads the medical records for a given patient, looking both at the structured and unstructured data. Then once it’s identified the key attributes for a given patient, Watson looks into a corpus of data that includes guidelines, textbooks, thousands of pages of journal articles, and published abstracts, while also relying on its expert training from doctors at Memorial Sloan Kettering over thousands of cases. Then it makes evidence-based treatment recommendations for the physician to consider in deciding how to treat a given patient.
Could you talk more about structured and unstructured data?
In healthcare, a lot of the most important information about a given patient exists in an unstructured narrative form in the electronic medical records. Doctors’ notes are highly variable from one to the next. When you look at those narrative documents, you can’t necessarily predict where in the given document a crucial piece of information might reside. Watson works by reading the text, understanding the content, making conceptual links between sentences and paragraphs, and ultimately is able to extract from that unstructured text the most important pieces of information. It’s able to do that from doctors’ notes, radiology reports, oncology reports, laboratory results, etc. It then compares that against these terrific knowledge sources that are constantly being updated.
How does Watson Clinical Trials Matching make finding matches easier?
There are a lot of barriers to clinical trial enrollment. Every clinical trial in which a cancer patient might participate carries with it a long and detailed list of eligibility criteria that must be met in order for a patient to be enrolled. The eligibility criteria includes things like the cancer type and stage, other co-morbid conditions, and medications that the patient might be on prior to therapy. A given physician who’s enrolling patients in clinical trials typically has tens or even hundreds of trials to consider. The lists of eligibility criteria are generally tens of items long, so that’s obviously a hard thing for any human to keep track of. Watson Clinical Trials Matching benefits physicians and patients in much the same way Watson for Oncology does. It reads the medical records, structured and unstructured data, and pulls out key attributes. It can then, in an automated fashion, compare the patient attributes to the eligibility criteria and help physicians make matches.
How is Watson offering increased personalization to doctors and patients?
I look at Watson as an engine for personalized medicine therapy because its first step is to extract all of the known, relevant attributes of a given patient that need to be considered in making treatment recommendations. As one looks to the future, I think there’s also the reality that there are many likely undiscovered attributes of given patients and given tumors that will impact treatment. I believe that Watson has the capacity to help us identify what some of these factors are.
Watson Oncology has a high concordance with doctors when it comes to treatment recommendations—above 80 or even 90 percent in most studies. Is the goal eventually to get to 100 percent concordance?
This 80, 90 percent that you cite, it is exactly what you say, the concordance between Watson and expert human physicians. But the studies were not intended to answer the question of, “Which is right?” In other words, if you get an 80 percent concordance number, it may be equally probable that Watson is right or the humans are right. If you take a group of well-trained doctors, even working in one hospital, give them a case, and ask for their treatment recommendations, I think, in many cases, you would find quite a low concordance rate between them. That doesn’t necessarily mean that those doctors are right or wrong. Sometimes, it means that there are multiple good answers.
How do you see cognitive computing developing in the future for healthcare? Do you see Watson eventually having the capacity to replace doctors in any way?
I don’t expect that Watson is going to replace doctors. That’s definitely not IBM’s intent, and I don’t think it’s realistic. I do think that cognitive computing in healthcare is going to become increasingly ubiquitous because of the reality that there is such a growing cognitive burden based on the rapid acceleration of knowledge in healthcare. I think that there are some things that cognitive computing systems do extraordinarily well. They can find patterns in large data sets that humans can’t find. They can search very, very quickly, they don’t need to sleep, they have unlimited storage, they’re often not subject to the same kind of cognitive biases that people are—that sort of thing. That said, there is and always will be, I think, a critical role for physicians to sit and look at patients eye-to-eye, to help patients weigh options, to be a compassionate voice, to help people explore their goals, fears, and hopes. The offloading of some mundane or routine tasks from physicians by cognitive computing systems will free them up to be more available for those strictly human types of interactions.
Watson Oncology focuses on treatment recommendations, but are there any other potential uses for Watson in the realm of healthcare?
We’re looking at a wide variety of different use cases for cognitive computing across multiple different areas of healthcare. It’s true that’s today’s Watson Oncology model brings our focus substantially on the provision of treatment recommendations for patients with diagnosed cancer. There’s no doubt that cognitive computing capabilities can have a role in, for example, prediction of disease. Patient subgroups who are at particularly high risk for an illness who might benefit from additional testing, or even prevention strategies. One of our major areas of focus that overlaps with oncology is in imaging. Watson Health Imaging is developing tools to improve the accuracy rate of tomography, for example. Way too many mammogram readings result in false positive or false negative results. Cognitive computing technology is typically the ability of Watson to interpret medical images based on training. We think it has real potential to improve the accuracy of these imaging-based screening studies.
Source:https://www.psfk.com/2017/07/ibm-watson-shows-us-machine-learning-transforming-cancer-care.html
Comments
Post a Comment